lama_dilated_int8

LaMa-Dilated INT8 - On-Device Image Inpainting Model

Model Name: LaMa-Dilated Model Type: Image editing / High resolution image in-painting / Image Inpainting Base Model: LaMa-Dilated (original implementation from advimman/lama) Input Resolution: 512×512 pixels (RGB image + 1-channel mask) Format: TFLite (INT8 quantized) / FP16 original Model Size: 44 MB (INT8 quantized) / 174 MB (FP16 original) Number of Parameters: 45.6M Tags: backbone, android Pipeline Tag: image-to-image Primary Compute Unit: NPU Precision: INT8

---

Download

File Size Link lama_int8.tflite 44 MB Download from Hugging Face

https://huggingface.co/Mateus24232u/lama_dilated_int8/resolve/main/lama_dilated_int8.tflite

SHA-256 Checksum: fb37b19e746f71c930fc42d1cc4f0b3582fbe003f8aff531212f2dfc240d88f7

https://huggingface.co/Mateus24232u/lama_dilated_int8/resolve/main/lama_dilated_int8.tflite.sha256

To verify integrity after download:

sha256sum lama_int8.tflite

Original FP16 model (174 MB): Qualcomm AI Hub

---

Model Description

LaMa-Dilated (Large Mask Inpainting) is a state-of-the-art image inpainting model designed to remove objects, fill holes, and repair damaged areas in images. This model is specifically optimized for on-device deployment with INT8 quantization, achieving fast inference on mobile NPUs.

Key Features

· Removes unwanted objects from images · Repairs scratches, cracks, and defects · Fills large masked areas with coherent texture · Preserves image structure and fine details · Works entirely offline after download

Inputs

Name Shape Format Description image [1, 512, 512, 3] RGB uint8 Input image to inpaint mask [1, 512, 512, 1] Binary uint8 (0 or 255) White where to inpaint, black elsewhere

Output

Name Shape Format Description output [1, 512, 512, 3] RGB uint8 Inpainted image with mask region filled

Performance

Device Chipset Inference Time Samsung S23 Ultra Snapdragon 8 Gen 2 88.6 ms (NPU) Pixel 8 Pro Tensor G3 125.3 ms (NPU) Xiaomi 13 Pro Snapdragon 8 Gen 2 91.2 ms (NPU)

Peak Memory Range: 3 - 133 MB (TFLite)

---

References

· Original LaMa paper: Resolution-robust Large Mask Inpainting with Fourier Convolutions · Original implementation: advimman/lama · Qualcomm AI Hub: LaMa-Dilated

---

License

Component License Original implementation Apache 2.0 Compiled TFLite asset (FP16) Qualcomm AI Hub Proprietary License INT8 quantized version Same as FP16 (derived work)

Neither license is GPL, making this model safe to use in open-source applications with permissive licenses. No conflict with FCam or any other open-source camera application.